Hello,
I want to implement a HTM using the paper “Continuous online sequence learning with an
unsupervised neural network model” but I’m a little confused about some terminology.
Did I understand correctly that:
- there can be multiple distal segments (and are in fact used in the simple test scenarios)
- each segment has multiple synapses
- to put the cell into the predictive state it is enough that a SINGLE segment to exceeding threshold
Is that correct ?
In " 3.3. HTM activation and learning rules" the paper is sometimes talking about a permanence matrix for synapses (which makes sense to me) and a permanence matrix for segments (which I don’t really understand).
Is this a mistake ?
If a cell (column) is correctly predicted it is clear what to do, but in case a cell (column) is wrongly or not at all predicted leaves some questions:
- when to create new segments (if at all) vs. adjust only permanences of synapses on existing segments (and which ones in case there are multiple)
- what exactly does “not matching segments” vs, “matching segments” mean if they could not exceed the threshold to put the cell into a predictive state ?
Hope somebody can help me out a little.
Cheers and thanks in advance 
